Mastermind recap
AIMM Session — July 3, 2025: Voice Bot in the Wild, GEO Discovery, and Avatar Archaeology
“Information is easy. Knowledge — sweat, frustration, pain, but ultimately transformation.” — Don Back, in the chat
30-Second Summary
A small but fired-up crew gathered the day before the US holiday. Dirk’s voice bot is out in the wild and producing delightful — and occasionally delusional — results. Lou is deep inside a legal AI build. Don unlocked a hidden avatar with AI-powered client analysis. Kasimir cracked a Facebook ads mystery. And the group tumbled into a genuinely meaty conversation about AI copyright, LLM SEO, and why running a frontier model locally still isn’t worth it.
1. Your Voice Bot Is Live — Now What?
Dirk’s 11 Labs agent is deployed. It already invented a salary figure for Dirk out of thin air when someone asked a question outside its knowledge base. The group’s playbook:
- Start with an MVP knowledge base — get the narrow use case working before adding internet access
- Instruct the bot to cite all sources
- Make the bot transparent about when it’s using web search: “A recent search of the Internet shows…”
- Curate which sites it searches — if you have 3 trusted industry sources, let it only search those
- Use the transcripts — what users ask that can’t be answered is your product roadmap
2. ChatGPT Is Recommending You — And You Don’t Know Why
Dirk got a cold LinkedIn DM from a Swiss CIO who found him through ChatGPT. The CIO reached out calling him “an authority” based purely on an AI recommendation. Best theory: consistent, contextually relevant use of ChatGPT for executive search created a data footprint the model learned from. A secondary theory: controlling Bing may have more influence on ChatGPT’s recommendations than Google SEO.
The real insight: There’s an emerging discipline — LLM SEO or GEO (Generative Engine Optimization) — and almost nobody in the practitioner community has cracked it yet.
Try This: Ask ChatGPT who the top 3 experts are in your specific niche. If you’re not there, that’s your roadmap.
3. AI Copyright: The Legal Iceberg
Bally brought receipts from an AI governance panel:
- Training on data = green light (at least in the US, for now). The Anthropic case confirmed that using data to train a model is not, in itself, infringement.
- How the data was sourced = the actual battle. AI companies are being pursued for where that training data came from — specifically, scraping millions of books from LibGen without authorization.
- EU vs. US divergence is real. The EU AI Act has teeth. Extraterritorial implications are significant.
4. Content Automation That Actually Runs on Autopilot
Dirk’s automated publishing stack:
- New candidate enters Zoho CRM with a 5-star rating → Make.com triggers → generates a LinkedIn post framed as a “growth opportunity” story → publishes simultaneously to LinkedIn and his website blog.
Kasimir’s content prompt tip:
“Top 0.1% expert in [field] — what is a contrarian and interesting point of view for [target audience]?”
Don’s content engine: 6-month LinkedIn content calendar built with ChatGPT using his ideal client avatar profile and 5 emotional themes.
5. Don’s Hidden Avatar — “Project Amy”
Don went back through his client database and asked ChatGPT: “What am I not seeing? What attributes do my most successful clients share?” The AI surfaced a psychographic pattern: the majority of his best clients are women in early-to-mid career, seeking or navigating a career path change, often with advanced degrees — but not necessarily PhDs. Relaxing the PhD constraint unlocked a much wider addressable market.
Try This: Drop your last 10–15 clients into ChatGPT. Ask: “What psychographic, demographic, and behavioral patterns do these people share that I might not have noticed? What constraint am I imposing that’s limiting my market?“
6. Local AI — Honest Numbers
| Setup | Speed | Cost | Verdict |
|---|---|---|---|
| Llama 3 70B (8-bit quantized) on M4 Mac, 24GB RAM | ~13 tokens/sec, ~30s think time | Hardware cost | Painful for business use |
| Llama 3 70B on RTX 6000 GPU ($8K card) | ~30 tokens/sec | $8,000 hardware | Still too slow |
| Llama 3 70B on Groq (cloud) | 200–700 tokens/sec | Pennies per call | Actually usable |
| Claude / GPT-4o (frontier, cloud) | Near-instant | Subscription/API | Best for business |
Resources Mentioned
- ElevenLabs — Voice agent platform Dirk is using
- Make.com — Lou’s recommended automation tool for the non-technical majority
- N8n (self-hosted) — Lou’s personal preference for automation with maximum control
- Groq — Cloud inference provider; runs Llama 3 70B at 200–700 tokens/second
- Pabbly Connect — Mentioned by Waldon as a Zapier alternative
- Michael Simmons — Free 4-part course available on his website